AWS adds GPT-5.6 tiers to Bedrock, with regional and routing constraints

AWS says GPT-5.6 Sol, Terra, and Luna are generally available through Amazon Bedrock, with pricing aligned to OpenAI's first-party rates and usage counting toward existing AWS commitments. For organizations already operating model workloads in AWS, the practical change is a new access path for the GPT-5.6 family inside Bedrock rather than a separate procurement or identity setup. AWS positions that path around its next-generation inference engine, in-region processing, and the controls surrounding a Bedrock deployment.

OpenAI describes Sol as the flagship reasoning tier, Terra as the lower-cost tier positioned against GPT-5.5, and Luna as the fastest and least expensive tier. AWS carries the same three-tier framing into Bedrock and recommends different workload shapes for each: deep multi-step work for Sol, general production tasks for Terra, and high-volume classification, summarization, routing, or real-time use for Luna. That makes the choice more than a model-name substitution, because an application can route work by capability and expected spend.

The Bedrock availability is not geographically uniform. AWS lists Sol in US East (N. Virginia) and US East (Ohio), while Terra and Luna are also listed for US West (Oregon). Teams that require Sol therefore need to account for its narrower listed regional footprint when designing data-residency, latency, and failover plans. AWS also says requests can remain in the AWS Region selected by the customer, which makes the region choice part of the deployment design rather than a purely operational detail.

Cost control is another reason the tiers may be used differently. OpenAI lists Sol at $5 per million input tokens and $30 per million output tokens, Terra at $2.50 and $15, and Luna at $1 and $6. Both OpenAI and AWS describe explicit prompt-cache breakpoints, a 90 percent discount for cached input, and a minimum 30-minute cache life. Those details favor agents with repeated system instructions, tool definitions, or reference files, but they do not remove the need to measure token mix and cache-hit behavior in the target workload.

The performance rationale should be read as provider-reported evidence, not as a neutral deployment benchmark. AWS explicitly attributes several Sol performance figures to OpenAI, while OpenAI's comparison tables combine published benchmark names with its own reported and internal evaluations. The practical next step for technical users is to test representative prompts, tool loops, failure handling, regional latency, and cache behavior on Bedrock before treating the catalog claims as application-level results. The sources establish the availability and interface direction; they do not establish that one tier is best for every workload.